PRIVACY PRESERVING FEDERATED LEARNING WITH INTEGRATED MALICIOUS CLIENT IDENTIFICATION

Abstract

A federated learning architecture that protects users' privacy and incorporates a malicious client detection system is proposed in this research. The goal of the design is to improve the safety and reliability of collaborative model training in remote settings. Data privacy is protected with federated learning (FL), which allows several clients to train a shared global model without revealing raw data. However, FL is still susceptible to model manipulation, poisoning assaults, and unreliable participation. In order to address these issues, a robust client evaluation module is integrated with safe aggregation and differential privacy techniques. This module uses statistical deviation analysis and adaptive trust score to detect odd updates. The method improves the model's stability, convergence speed, and overall performance by identifying and resolving malicious or faulty client contributions prior to the global model aggregation. Experimental validation proves that the integrated strategy achieves a balance between privacy protection, attack resistance, and model correctness. Its security features make it an ideal choice for decentralized smart system, healthcare, and financial applications.